In order to make full use of the correlation and complementarity between multimodal medical images,so as to accurately segment brain tumor regions and evaluate the prognostic effect,we proposed a multimodal 3D segmentation model for brain tumors based on residual siamese network.Firstly,the residual siamese coding was used to excavate the relational semantic information be-tween the different modal data,and a cascade structure was added between the coding paths to optimize the information interaction be-tween levels.Additionally,a multi-scale pixel attention fusion block was proposed to obtain weighted fusion features,promote the ex-change of complementary information among modalities.Finally,in the decoding stage,skip connections and attention gates based on the residual siamese encoding structure were used to guide the model to focus on information relevant to tumor segmentation,thereby improving segmentation performance.The experiment was verified on the BraTS 2021 dataset,the average Dice coefficients in the whole tumor,tumor core and enhanced tumor regions reached 0.928,0.914 and 0.879,respectively.This research is expected to offer a new method for the early diagnosis of clinical brain diseases.
维普
万方数据